biglasso

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.width=6, fig.height=4
)

1 User Guide

1.1 Small Data

1.1.1 Standar Lasso

library(biglasso)

data(colon)
X <- colon$X
y <- colon$y
dim(X)
X[1:5, 1:5]
## convert X to a big.matrix object
## X.bm is a pointer to the data matrix
X.bm <- as.big.matrix(X)
str(X.bm)
dim(X.bm)
X.bm[1:5, 1:5]
## same results as X[1:5, 1:5]
## fit entire solution path, using our newly proposed "Adaptive" screening rule (default)
fit <- biglasso(X.bm, y)
plot(fit)

1.1.2 Cross-Validation

## 10-fold cross-valiation in parallel
cvfit <- tryCatch(
         {
                cv.biglasso(X.bm, y, seed = 1234, nfolds = 10, ncores = 4)
         },
         error = function(cond) {
                cv.biglasso(X.bm, y, seed = 1234, nfolds = 10, ncores = 2)
         }
)

After cross-validation, a few things we can do:

par(mfrow = c(2, 2), mar = c(3.5, 3.5, 3, 1) ,mgp = c(2.5, 0.5, 0))
plot(cvfit, type = "all")
summary(cvfit)
coef(cvfit)[which(coef(cvfit) != 0),]

1.1.3 Logistic Regression

data(Heart)
X <- Heart$X
y <- Heart$y
X.bm <- as.big.matrix(X)
fit <- biglasso(X.bm, y, family = "binomial")
plot(fit)

1.1.4 Cox Regression

library(survival)
X <- heart[,4:7]
y <- Surv(heart$stop - heart$start, heart$event)
X.bm <- as.big.matrix(X)
fit <- biglasso(X.bm, y, family = "cox")
plot(fit)

1.1.5 Multiple responses Linear Regression

set.seed(10101)
n=300; p=300; m=5; s=10; b=1
x = matrix(rnorm(n * p), n, p)
beta = matrix(seq(from=-b,to=b,length.out=s*m),s,m)
y = x[,1:s] %*% beta + matrix(rnorm(n*m,0,1),n,m)
x.bm = as.big.matrix(x)
fit = biglasso(x.bm, y, family = "mgaussian")
plot(fit)

1.2 Big Data

When the raw data file is very large, it's better to convert the raw data file into a file-backed big.matrix by using a file cache. We can call function setupX, which reads the raw data file and creates a backing file (.bin) and a descriptor file (.desc) for the raw data matrix:

## The data has 1000 observations and 5,000 features.
## Much larger data can be handled in the same way.
## 10 of the features has non-zero coefficients
if(!file.exists('BigX.bin')) {
  X <- matrix(rnorm(1000 * 5000), 1000, 5000)
  beta <- c(-5:5)
  y <- as.numeric(X[,1:11] %*% beta)
  write.csv(X, "BigX.csv", row.names = F)
  write.csv(y, "y.csv", row.names = F)
  ## Pretend the data is stored in the ~90MB .csv file and is too large to fit into memory
  X.bm <- setupX("BigX.csv", header = T)
}

It's important to note that the above operation is just one-time execution. Once done, the data can always be retrieved seamlessly by attaching its descriptor file (.desc) in any new R session:

rm(list = c("X", "X.bm", "y")) # Pretend starting a new session
X.bm <- attach.big.matrix("BigX.desc")
y <- read.csv("y.csv")[,1]

This is very appealing for big data analysis in that we don't need to "read" the raw data again in a R session, which would be very time-consuming. The code below again fits a lasso-penalized linear model, and runs 10-fold cross-validation:

system.time({fit <- biglasso(X.bm, y)})
plot(fit)
# 10-fold cross validation in parallel
tryCatch(
    {
        system.time({cvfit <- cv.biglasso(X.bm, y, seed = 1234, ncores = 4, nfolds = 10)})
    },
    error = function(cond) {
        system.time({cvfit <- cv.biglasso(X.bm, y, seed = 1234, ncores = 2, nfolds = 10)})
    }
)
par(mfrow = c(2, 2), mar = c(3.5, 3.5, 3, 1), mgp = c(2.5, 0.5, 0))
plot(cvfit, type = "all")

2 Useful Reference



Try the biglasso package in your browser

Any scripts or data that you put into this service are public.

biglasso documentation built on March 18, 2022, 5:47 p.m.